An intelligent identification method and system for algal bloom risk remote sensing

Abstract

The present application belongs to the technical field of water ecological risk early warning, and provides an algal bloom risk remote sensing intelligent identification method and system, remote sensing image data, meteorological data and water quality data of a target area are acquired and preprocessed; frequency domain feature extraction is performed to generate a three-dimensional frequency domain feature vector; a multi-task deep learning network is used to fuse the global feature representation obtained by modeling and the generated three-dimensional frequency domain feature vector to obtain a feature map; based on the feature map, an algal bloom binary segmentation probability map and a continuous value distribution map of chlorophyll a concentration are obtained; a semantic embedding head is used to generate an image semantic embedding vector, a pre-trained knowledge graph is used to perform semantic alignment on the image semantic embedding vector to generate research and judgment information text; the research and judgment information text, the algal bloom binary segmentation probability map and the continuous value distribution map of chlorophyll a concentration are cross-validated to obtain a final algal bloom risk judgment result. The present application realizes algal bloom risk remote sensing intelligent identification.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of water ecological risk early warning, and particularly relates to an algal bloom risk remote sensing intelligent identification method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Algal bloom is a frequent water ecological disaster in lakes, reservoirs, estuaries and nearshore sea areas. Its outbreak has seriously damaged the balance of the water ecological system, and threatened the safety of drinking water, aquaculture industry and public health.

[0004] Machine learning and deep learning classification methods mostly use algorithms such as support vector machine, random forest or convolutional neural network to automatically learn features from remote sensing images. Compared with the traditional index threshold method, although the recognition accuracy is improved, there are still significant limitations.

[0005] Firstly, the existing methods are mostly single modal data processing, lacking effective fusion mechanism for meteorological, water quality and other multi-source heterogeneous data. The digital processing ability of deep mining of complex correlations among remote sensing, meteorological, water quality and other heterogeneous data is insufficient, and the key environmental driving factors affecting the growth and decline process of algal bloom are not fully utilized, which limits the in-depth analysis of the causes of algal bloom.

[0006] Secondly, the model output is mostly abstract algal bloom distribution binary graph or classification graph, and the structured semantic information representation ability of algal bloom risk level, spatial form feature, etc. is insufficient, and the explainability needs to be improved.

[0007] In addition, the intelligent analysis level of remote sensing extraction results is not good, and the automatic generation degree of comprehensive research and judgment report such as algal bloom situation, environmental causes, risk level and prevention and control suggestions is not high, which still needs to rely on the manual interpretation and analysis of interpreters, and the experience level is required to be high. The business and real-time ability of algal bloom risk prediction and early warning still needs to be improved. SUMMARY

[0008] In order to solve the above problems, the present application provides an algal bloom risk remote sensing intelligent identification method and system, which can improve the algal bloom recognition accuracy by using frequency domain feature extraction model, and realize intelligent interpretation through image-semantic cross-modal alignment.

[0009] According to some embodiments, the present application adopts the following technical solutions:

[0010] An algal bloom risk remote sensing intelligent identification method, comprising the following steps:

[0011] Obtaining remote sensing image data, meteorological data and water quality data of a target area;

[0012] The obtained remote sensing image data, meteorological data and water quality data are respectively preprocessed;

[0013] The preprocessed remote sensing image data is subjected to frequency domain feature extraction, a set of band gray image blocks are selected, fast Fourier transform is performed to obtain complex spectrum, power spectral density is calculated, the power spectral density spectrum is divided into frequency regions of different energy levels and the average energy is calculated, and a three-dimensional frequency domain feature vector is generated to represent the spatial texture pattern of the algal bloom;

[0014] A multi-source data cube is constructed, which is stacked with the processed remote sensing image data, meteorological data and water quality data;

[0015] Using a multi-task deep learning network, the multi-source data cube is sequentially subjected to spatial feature extraction, serialization and global feature modeling, the global feature representation obtained by modeling and the generated three-dimensional frequency domain feature vector are fused to obtain a feature map;

[0016] Based on the feature map, an algal bloom binary segmentation probability map is obtained using a segmentation head, and a continuous value distribution map of chlorophyll a concentration is obtained using a hierarchical head;

[0017] Based on the global feature representation, an image semantic embedding vector is generated using a semantic embedding head, and the image semantic embedding vector is semantically aligned using a pre-trained knowledge graph to generate a research and judgment information text;

[0018] The research and judgment information text, the algal bloom binary segmentation probability map and the continuous value distribution map of chlorophyll a concentration are cross-validated to obtain the final algal bloom risk judgment result.

[0019] As an optional implementation, the process of preprocessing the obtained remote sensing image data includes: performing radiation calibration on the remote sensing image data, converting the digital quantization value into the top-of-atmosphere radiance, and performing atmospheric correction on the remote sensing image data to obtain the water body remote sensing reflectance; in the water body area, the normalized difference chlorophyll index and the phytoplankton index are calculated.

[0020] As an optional implementation, the process of preprocessing the obtained meteorological data includes: interpolating the meteorological data to the same spatial grid as the remote sensing image, calculating the daily average value or the cumulative value of the previous day matched with the transit time of the remote sensing image; the wind speed vector is decomposed into east and north components.

[0021] As an optional implementation, the process of preprocessing the obtained water quality data includes: the water quality data includes the concentrations of chlorophyll a, total nitrogen and total phosphorus, the quartile range method is used to remove abnormal measurement values, if the density of monitoring sites is less than a set value, inverse distance weighting or Kriging interpolation is used to generate a spatial distribution map, the measurement value closest to the date of the remote sensing image is taken, and if the dates do not match, linear interpolation is performed.

[0022] As an alternative embodiment, the process of dividing the power spectral density spectrum into different energy level frequency regions and calculating the average energy includes: dividing the power spectral density spectrum into three fan-shaped regions of low frequency, medium frequency and high frequency in polar coordinates, and calculating the average energy value of each region, wherein the low frequency fan-shaped region is used to represent the components in the image with a speed change less than a set value, corresponding to large area, uniform distribution of algae;

[0023] The medium frequency fan-shaped region is used to represent the components in the image with a speed change within a set range, corresponding to filamentous, strip and block structure of algae;

[0024] The high frequency fan-shaped region is used to represent the components in the image with a speed change greater than a predetermined value, corresponding to fragmented distribution, algae boundary contour or noise;

[0025] Based on the average energy value calculated in each fan-shaped region, a three-dimensional frequency domain feature vector is constructed as a quantitative description of the spatial distribution pattern of algae.

[0026] As an alternative embodiment, the process of constructing a multi-source data cube includes: the shape of the multi-source data cube is [Batch_size, Channels, Height, Width], wherein Batch_size is the number of samples input to the model for processing at a time, Channels dimensionally stacks the features of remote sensing image data, meteorological data and water quality data, Height is the spatial height dimension after the remote sensing image is cropped, and Width is the spatial width dimension after the remote sensing image is cropped.

[0027] As an alternative embodiment, the multi-task deep learning network includes an encoder for preliminary spatial feature extraction of the multi-source data cube, dividing the extracted feature map into a plurality of fixed size image blocks according to spatial position, flattening the features of each image block and mapping them to a uniform dimension through linear projection to form a group of serialized feature vectors, adding a learnable classification token feature at the front end of the serialized feature vector for aggregating global information, and adding the serialized feature vector after adding through a multi-head self-attention mechanism to realize global feature modeling and obtain a representation with global dependency relationship;

[0028] And project the three-dimensional frequency domain feature vector through a fully connected layer to the same dimension as the global feature, and add and fuse it with the classification token feature element by element to enhance its feature representation.

[0029] As an alternative implementation, the multi-task deep learning network comprises a decoder, the decoder comprising a segmentation head, a hierarchical head and a semantic embedding head, the segmentation head being configured to receive the feature map output by the encoder, output a bloom binary segmentation probability map through convolution and a Sigmoid activation function, so as to depict the spatial distribution range and boundary of the bloom, thereby supporting bloom distribution mapping and area calculation;

[0030] The hierarchical head is configured to receive the feature map output by the encoder, obtain a continuous value distribution map of chlorophyll a concentration through convolution and linear activation function, and discretize the continuous value distribution map into a risk level map, so as to perform bloom grade division and risk warning;

[0031] The semantic embedding head takes the global semantic vector output by the encoder as input, generates an image semantic embedding vector through projection transformation, the image semantic embedding vector encoding the overall bloom condition of the image block and its environmental context information, and aligning with the structured semantic label and the ecological knowledge graph, so as to realize semantic analysis of the remote sensing image result.

[0032] As an alternative implementation, the training process of the knowledge graph comprises:

[0033] A structured semantic label library is constructed, and a structured label containing bloom state, dominant species, density level and environmental condition is generated for each training sample, and converted into a natural language description sequence;

[0034] A large language model is used to encode the text description into a feature vector, and a loss function is used for training to reduce the distance between matching image-text pairs and increase the distance between non-matching pairs, so as to realize cross-modal semantic alignment;

[0035] For each structured label, a corresponding text embedding vector is output by the trained large language model, and based on the structured label and the text embedding vector, a knowledge graph is formed.

[0036] An algal bloom risk remote sensing intelligent identification system, comprising:

[0037] A data acquisition module configured to acquire remote sensing image data, meteorological data and water quality data of a target area;

[0038] A preprocessing module configured to preprocess the acquired remote sensing image data, meteorological data and water quality data respectively;

[0039] A feature extraction module configured to perform frequency domain feature extraction on the preprocessed remote sensing image data, select a set of band gray image blocks, perform fast Fourier transform to obtain a complex spectrum, calculate the power spectral density, divide the power spectral density spectrum into different energy level frequency regions and calculate the average energy, and generate a three-dimensional frequency domain feature vector to represent the spatial texture pattern of the bloom;

[0040] a data stacking module configured to construct a multi-source data cube stacked with processed remote sensing image data, meteorological data and water quality data;

[0041] a multi-task learning module configured to sequentially perform spatial feature extraction, serialization and global feature modeling on the multi-source data cube by using a multi-task deep learning network, fuse the global feature representation obtained by modeling and the generated three-dimensional frequency domain feature vector to obtain a feature map;

[0042] a preliminary judgment module configured to obtain an algal bloom binary segmentation probability map by using a segmentation head and obtain a continuous value distribution map of chlorophyll a concentration by using a hierarchical head based on the feature map;

[0043] a semantic extraction module configured to generate an image semantic embedding vector by using a semantic embedding head based on the global feature representation, perform semantic alignment on the image semantic embedding vector by using a pre-trained knowledge graph, and generate a judgment information text;

[0044] a cross-validation module configured to cross-validate the judgment information text, the algal bloom binary segmentation probability map and the continuous value distribution map of chlorophyll a concentration to obtain a final algal bloom risk judgment result.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] The present application deeply fuses the spatial domain visual features and the frequency domain features of multi-scale algal blooms, enhances the perception ability of the frequency domain information enhancement model to texture and periodic distribution patterns, and significantly improves the accuracy and boundary clarity of algal bloom recognition and segmentation. On this basis, the model realizes the leap from image recognition to semantic interpretation through the parallel segmentation head, hierarchical head and semantic embedding head, and can simultaneously output high-precision algal bloom distribution maps, chlorophyll a concentration grade maps and embedding vectors encoding global semantics.

[0047] The present application realizes the integrated output of graphs, spectra and texts, provides a full-chain intelligent solution integrating precise monitoring, risk early warning and causal analysis for algal bloom supervision of lakes, reservoirs, estuaries and nearshore sea areas, and improves the automation, interpretability and decision support level of environmental supervision.

[0048] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are used for explanation. BRIEF DESCRIPTION OF DRAWINGS

[0049] The drawings accompanying the specification of the present application form part of the present application and are used to provide a further understanding of the present application, the illustrative embodiments of the present application and the description thereof serve to explain the present application and do not constitute an improper limitation of the present application.

[0050] Figure 1 Figure 1 is a schematic diagram of a method for intelligent identification of algal bloom risk by remote sensing according to an embodiment;

[0051] Figure 2 Figure 2 is a schematic diagram of algal bloom aggregation characteristics of a remote sensing image in a set waveband according to an embodiment;

[0052] Figure 3 Figure 3 is a schematic diagram of low-frequency, medium-frequency and high-frequency fan-shaped region division according to an embodiment;

[0053] Figure 4 Figure 4 is a binary segmentation map of algal bloom according to an embodiment. DETAILED DESCRIPTION

[0054] The application will be further described below with reference to the drawings and embodiments.

[0055] It should be noted that the following detailed description is illustrative only and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0056] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be further understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of a feature, step, operation, device, component and / or combinations thereof.

[0057] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0058] Embodiment One

[0059] An intelligent identification method of algal bloom risk by remote sensing, as shown in Figure 1, includes the following steps: Figure 1

[0060] Step 1, data collection and preprocessing

[0061] The present embodiment collects multi-source heterogeneous data of the research area for standardized processing, so that it is aligned in time and space scales to form an input vector available for the model.

[0062] The multi-source heterogeneous data includes remote sensing image data, meteorological data and water quality data.

[0063] The preprocessing of each part includes:

[0064] ​Remote sensing image data preprocessing: pre-processing of remote sensing data (GOCI, Sentinel, etc.) such as radiometric calibration, atmospheric correction, etc.

[0065] Among them, radiometric calibration: convert digital quantization value (DN) ) to top-of-atmosphere radiance (TOA Radiance).

[0066] . Among them, is the gain, is the offset value.

[0067] Atmospheric correction: atmospheric correction of remote sensing images using Acolite to obtain water body remote sensing reflectance (Rrs). ;

[0068] Among them, is the exo-atmospheric radiance (unit: W·m -2 ·nm -1 ·sr -1 ), is the exo-atmospheric solar irradiance (unit: W·m -2 ·nm -1 ), is the solar zenith angle.

[0069] Feature calculation: in the water area, calculate the normalized difference chlorophyll index :

[0070] ;

[0071] Among them, is the near-infrared band remote sensing reflectance, is the red band remote sensing reflectance.

[0072] Phytoplankton index :

[0073] ;

[0074] Among them, is the near-infrared band remote sensing reflectance, is the red band remote sensing reflectance. are the center wavelengths (unit: nm) of the near-infrared, red and short-wave infrared bands, respectively.

[0075] Weather data preprocessing:

[0076] Analyze data ERA5 or site data, including sea surface temperature (SST), wind speed (WS), wind direction (WD) and precipitation (Precip).

[0077] Interpolate weather data to the same spatial grid as remote sensing imagery (e.g. Kriging interpolation). Calculate daily mean or previous day cumulative values that match the remote sensing data overpass time. Decompose wind speed vector into U (eastward component) and V (northward component).

[0078] Water quality data preprocessing:

[0079] Water quality data mainly includes chlorophyll-a (Chl-a), total nitrogen (TN), total phosphorus (TP) concentrations measured by manual or site monitoring. IQR (interquartile range) method is used to remove abnormal measurement values. In the case of sparse monitoring sites, Kriging interpolation is used to generate spatial distribution map. Take the measurement value closest to the date of remote sensing image, if the date does not match, linear interpolation is performed.

[0080] Use Fourier transform-based Transformer module to extract multi-scale spatio-temporal features of remote sensing images. Fuse water color index (such as normalized difference chlorophyll index NDCI and phytoplankton index FAI) with ecological prior indicators.

[0081] First, the following processing is performed:

[0082] (1) Frequency domain feature extraction

[0083] Algal blooms in water form specific spatial texture features (such as filamentous, sheet-like, and punctate), which have unique responses in spatial frequency. This embodiment converts the image to the frequency domain by Fast Fourier Transform (FFT) and calculates the Power Spectral Density (PSD), thereby extracting the frequency domain feature representing the spatial distribution pattern of algal blooms. This feature is complementary to the spectral feature, which is jointly input into the downstream network to improve the accuracy of algal bloom type discrimination and boundary recognition.

[0084] As shown in Figure 2 , the input is a gray image block I ∈ R H×W of remote sensing image red or near-infrared band (e.g. 745 nm). I is a two-dimensional matrix belonging to the real number domain, with a shape of height H pixels by width W pixels.

[0085] Transformation and calculation:

[0086] Fast Fourier Transform (FFT) is performed on the input I: F = FFT(I). The spatial domain image is transformed into the frequency domain to obtain a complex spectrum F.

[0087] Calculate the power spectral density (PSD): P = |F| 2 . Take the square of the modulus of the spectrum to get the distribution of signal energy in the frequency domain.

[0088] As shown in Figure 3Feature engineering: The 2D PSD image is divided into three sectors in polar coordinates, and the average energy value of each sector is calculated.

[0089] Low-frequency energy: represents the slowly changing component in the image, corresponding to large-area and uniformly distributed algae.

[0090] Mid-frequency energy: represents the medium-speed changing component in the image, corresponding to filamentous, strip, and block structure of algae.

[0091] High-frequency energy: represents the rapidly changing component in the image, corresponding to fragmented distribution, algae boundary contour, or noise.

[0092] Output: Generate a frequency domain feature vector F_freq, which is a 3-dimensional real number vector, where F_freq = [E_low, E_mid, E_high] T ∈ R 3 , as a quantitative description of the spatial distribution pattern of algae.

[0093] (2) Multi-task deep learning network construction

[0094] The network adopts a shared encoder-specialized decoder head architecture to balance feature sharing and task-specific processing. The overall architecture is a shared encoder plus three specialized decoder heads.

[0095] Input: X is a multi-source data cube, shaped as [Batch_size, Channels, Height, Width]. Among them, Batch_size is the number of samples (image blocks) input to the model for processing at a time, Channels dimensionally stacks various optical and environmental features such as image band reflectance, NDCI, FAI, SST, U, V, Precip, Chl-a, TN, TP, etc., Height is the spatial height dimension of the cropped remote sensing image, and Width is the spatial width dimension of the cropped remote sensing image. F_freq is a frequency domain feature vector [Batch_size, 3], where the 3 components represent the low, medium, and high frequency energy features extracted from the PSD, which are input as a separate global feature vector in the later stage of the network and fused with the encoder output.

[0096] Encoder (Encoder):

[0097] Feature preliminary extraction: First, a 3x3 two-dimensional convolutional neural network (2D-CNN) is used to preliminarily extract spatial features from X.

[0098] Sequence and global modeling: The feature maps extracted by CNN are divided into several fixed-size image patches according to the spatial position. The features of each patch are flattened and mapped to a unified dimension D through linear projection to form a set of sequence feature vectors. A learnable [CLS] token (Classification Token) is added at the front of the sequence to aggregate global information. The sequence is input into the Vision Transformer (ViT) encoder to realize global feature modeling through the multi-head self-attention mechanism, thereby obtaining a representation with global dependency.

[0099] In the Transformer, the formula for calculating the weight and the corresponding output is:

[0100] ;

[0101] The core multi-head self-attention (MSA) mechanism formula is:

[0102] ;

[0103] ;

[0104] Where Q, K, and V are the query, key, and value matrices obtained by linear transformation of the input sequence; h is the number of attention heads; d k is the key vector dimension; W i Q , W i K , and W i V are the linear mapping weight matrices of each attention head; W O is the learnable weight matrix of the output projection, which is used to fuse the multi-head output and restore it to the unified feature space dimension of the model.

[0105] Frequency domain feature fusion: Project F_freq through a fully connected layer to the same dimension D as the ViT output feature, and add it element by element with the [CLS] token feature output by ViT to enhance its feature representation.

[0106] Decoding head (Heads):

[0107] First, the segmentation head (Segmentation Head) takes the feature map output by the encoder as input, and outputs a binary algae bloom segmentation probability map through convolution and Sigmoid activation function, as shown in Figure 4 , which is used to depict the spatial distribution range and boundary of algae bloom, thereby supporting algae bloom distribution map drawing and area calculation.

[0108] Secondly, the Grading Head also receives the feature map output by the encoder, and after convolution and linear activation functions, it obtains a continuous value distribution map of chlorophyll a concentration, which can be discretized into a risk level map according to industry standards for algal bloom classification and risk warning.

[0109] ;

[0110] For the semantic embedding head, the feature representation of the [CLS] token output by the encoder (i.e., the global semantic vector aggregated by the encoder) is used as input, and a high-level semantic embedding vector is generated through projection transformation. This embedding vector not only encodes the overall algal bloom status of the image patch and its environmental context information, but can also be aligned with structured semantic labels and ecological knowledge graphs, thereby realizing semantic parsing of remote sensing image results, supporting the automatic generation of comprehensive assessment reports, and significantly enhancing the interpretability and digital application potential of the results.

[0111] Segmentation Head: Consists of a 1×1 convolutional layer and a sigmoid activation function. It receives the ViT output feature map (reconstructed as [B, D, H', W']) and outputs a binary segmentation probability map Y_seg ∈ [0,1]. H×W Where B (Batch Size) represents the number of independent samples (i.e., image patches) processed simultaneously in a single forward propagation. D is the Feature Dimension, representing the length of the feature vector at each spatial location (a point on the feature map) output by the ViT encoder. It represents the richness of high-level semantic information learned by the model. H' and W' are the height and width of the feature map, respectively. Since ViT typically segments the input image into non-overlapping image patches for processing, its output feature map will be smaller in spatial size than the original input image. H' and W' are the spatial dimensions (in pixels) of this downsampled feature map. H and W are the height and width of the original input image, respectively. The ultimate goal of the segmentation head is to output a probability map that perfectly corresponds to the spatial resolution of the original input image. Grading Head: Consists of a 1×1 convolutional layer + linear activation function. It receives the same feature map and outputs a continuous chlorophyll a concentration map Y_grade ∈ R. H×W .

[0112] The semantic embedding head consists of two fully connected layers, a ReLU activation function, and L2 normalization. It further transforms the [CLS] token features output by ViT to generate a high-level semantic embedding vector E. img ∈ R dThis vector encodes global semantic information of the image patch (such as the overall algal bloom status and environmental context information).

[0113] (3) Cross-modal semantic alignment

[0114] Algal bloom phenomenon E, which interprets images based on large language models (such as BERT model). img "Analysis of information texts described in human language E" text This directly correlates image and text data, achieving spatial alignment within the same vector. For example, through contrastive learning, the semantic information "filamentous cyanobacteria, high concentration, medium area, weak wind environment" identified from the image is automatically matched and aligned with its perfectly matching text description "filamentous cyanobacteria, high concentration, medium coverage area, weak wind conditions" in the vector space.

[0115] Key inputs for automatic generation of analysis reports: During the reasoning phase, E img It can be used as a "search key" to query structured semantic tag libraries or knowledge graphs. The model can then generate semantically parsed content based on this vector, such as: "The study area contains cyanobacterial blooms, distributed in filaments, covering an area of ​​approximately 120 km²." 2 The environmental conditions are high temperature, eutrophication, and weak wind, with a high risk level for outbreaks. This is the bridge between "numerical prediction / distribution map" and "natural language report." Unlike simply outputting a segmentation map, the semantic embedding head gives the model "semantic interpretation capabilities," enabling it to answer questions directly needed for business applications, such as "what kind of algae," "under what conditions does it occur," and "what is the risk level."

[0116] Loss function:

[0117] ;

[0118] α, β, and γ are hyperparameters used to balance the importance of each task. seg and GT seg Y represents the binary segmentation probability map of algal blooms output by the model and the true segmentation label, respectively. grade and GT grade These represent the continuous chlorophyll a concentration output by the model and the actual concentration, respectively. E img and E text These represent the image semantic embedding vector and the text semantic embedding vector, respectively.

[0119] Cross-modal semantic alignment and knowledge fusion: This step endows the model with "cognitive" capabilities. Through contrastive learning, it achieves cross-modal alignment of image features and semantic features, linking visual features with human language and knowledge.

[0120] It mainly includes the following three key parts:

[0121] (1) Constructing a structured semantic tag library:

[0122] For each training image patch, manual annotation or structured labels extracted from reports are used, for example:

[0123] {“Algal bloom status”: “Present”; “Main type”: “Cyanobacteria”; “Concentration level”: “Medium”; “Area”: “>100km”; 2 Spatial morphology: filamentous; Water temperature: high; Nutrient status: eutrophic; Wind speed: low

[0124] In remote sensing image samples, these labels are used to characterize typical features of algal blooms, such as state, species, morphology, concentration, and environmental driving factors. Their structured label representation is as follows:

[0125] {

[0126] # Algal bloom states include "existence" and "non-existence".

[0127] "Algal bloom state": "Exists"

[0128] The main types of algae include "cyanobacteria", "diatoms", "dinoflagellates", "mixed", and "unknown".

[0129] "Main type": "Cyanobacteria"

[0130] # Concentration levels are divided into "oligotrophic", "mesotrophic", and "eutrophic".

[0131] "Concentration Level": "Enriched"

[0132] # Algal bloom areas are categorized as "<10 km²" 2 10-100 km 2 "、 100 km" 2 "、">100 km 2 ",

[0133] The spatial morphology of algal blooms mainly includes "sheet-like", "filamentous", and "dot-like".

[0134] "Spatial morphology": "filamentous"

[0135] # Water temperature is categorized as "low", "normal", and "high" (relative to multi-year averages).

[0136] "Water Temperature Condition": "High",

[0137] # Wind speed is divided into "low", "medium" and "high" (e.g., wind speed <3m / s is low, wind speed between 3 and 6m / s is medium, and wind speed >6m / s is high).

[0138] Wind speed conditions: Low.

[0139] }

[0140] During batch processing, these text sequences are stored as a two-dimensional matrix (or tensor) with shape [Num_Samples, Max_Sequence_Length], where each element is an integer index (token ID) representing a word or subword. This is then input into the BERT model for encoding into text embedding vectors. A pre-trained BERT model is used as the text encoder. The processing flow is as follows:

[0141] Input Layer:

[0142] Input the generated text sequence into the model.

[0143] The model first performs tokenization, breaking down sentences into subwords or tokens from the vocabulary.

[0144] The model adds a special [CLS] marker at the beginning of the sequence, and its final hidden state will be used as the aggregate representation of the entire sequence.

[0145] The model generates a corresponding embedding vector for each token, which is formed by adding the token embedding, segment embedding, and position embedding.

[0146] Deep Encoding:

[0147] The embedded sequence is passed through multiple Transformer encoder layers of BERT.

[0148] Each layer uses a self-attention mechanism to allow each token in the sequence to interact with all other tokens, thereby dynamically computing its representation based on the context. For example, the model learns that "High" or "high" has different meanings when describing temperature and when describing concentration.

[0149] After multiple layers of computation, each input token is transformed into a contextualized vector representation.

[0150] Output Layer:

[0151] Take the hidden state corresponding to the [CLS] tag output by the last Transformer layer as the summary vector of the entire input text sequence.

[0152] Optionally, a projection layer (usually a fully connected layer) can be appended after this vector to transform its dimensions to match those of the image embedding vector E. img Same dimension d.

[0153] Output:

[0154] Finally, for each structured semantic tag, the text encoder outputs a corresponding text embedding vector E. text ∈ R d .

[0155] This vector is a dense, low-dimensional numerical representation that compresses and encodes all the semantic information (algal bloom state, type, environmental conditions, etc.) in the original text, and can be used for subsequent contrastive learning computations.

[0156] (2) Comparative learning and alignment:

[0157] Objective: To bring paired image embeddings E closer together in a unified high-dimensional space. img and text embedding E text The distance between the samples increases, thus increasing the distance between the unpaired samples.

[0158] Implementation: Using the InfoNCE Loss function:

[0159] ;

[0160] in, This represents the contrastive learning loss value. is the cosine similarity function used to calculate the similarity between quantities u and v. img It is an image semantic embedding vector generated by the semantic embedding head. Is it related to the current image E? img The embedding vector of the correctly matched text description (positive sample). τ is a temperature hyperparameter (τ > 0) used to adjust the sharpness of the similarity distribution. Smaller τ values ​​make the model more sensitive to hard negative samples (i.e., negative samples with high similarity to positive samples). N is the size of the training batch, i.e., the total number of image-text pairs contained in a batch. The summation term in the denominator covers all text embeddings within the batch, including 1 positive sample and N-1 negative samples.

[0161] (3) Ecological knowledge graph query:

[0162] The system has a built-in knowledge graph that uses triples (subject-relation-object) to represent the rules of algal bloom ecology.

[0163] Rule Example 1:

[0164] If (water temperature > 25 °C) and (TN / TP ratio > 16) and (wind speed < 3 m / s);

[0165] The output will be (algal bloom risk = high, dominant group = cyanobacteria, morphology = filamentous).

[0166] Rule Example 2:

[0167] If (transparency < 1 m) and (chlorophyll a > 20 µg / L) and (flow rate < 0.1 m / s);

[0168] The output is (algal bloom risk = high, spatial pattern = blocky cover).

[0169] During the inference phase, the model automatically queries the knowledge graph, traversing all IF-THEN rules in the knowledge base. It matches the current environmental conditions with the rules, calculating the matching degree as (number of satisfied preconditions) / (total number of preconditions). For rules with a matching degree > 0.7, the conclusion is aggregated according to weight: Risk Level = Σ(Rule Weight × Matching Degree × Rule Conclusion). The inference results from the knowledge graph are cross-validated with the chlorophyll a concentration map output by the hierarchical head and the area output by the segmentation head in the model. When the graph predicts "high risk" and remote sensing also detects "high concentration, large area," the reliability of the system's early warning is greatly improved. This generates factual conclusions for subsequent report generation and result verification.

[0170] This embodiment deeply integrates the spatial domain visual features and frequency domain features of multi-scale algal blooms. By utilizing frequency domain information to enhance the model's perception of texture and periodic distribution patterns, it can significantly improve the accuracy and boundary clarity of algal bloom recognition and segmentation. This embodiment can clearly identify algal blooms and simultaneously output high-precision algal bloom distribution maps, making the recognition results visible.

[0171] Example 2

[0172] A remote sensing intelligent identification system for algal bloom risk includes:

[0173] The data acquisition module is configured to acquire remote sensing image data, meteorological data, and water quality data of the target area;

[0174] The preprocessing module is configured to preprocess the acquired remote sensing image data, meteorological data, and water quality data respectively.

[0175] The feature extraction module is configured to extract frequency domain features from the preprocessed remote sensing image data, select grayscale image blocks of a set band, perform fast Fourier transform to obtain complex spectrum, calculate power spectral density, divide the power spectral density spectrum into frequency regions of different energy levels and calculate average energy, and generate a three-dimensional frequency domain feature vector to characterize the spatial texture pattern of algal blooms.

[0176] The data stacking module is configured to construct a multi-source data cube, which stacks processed remote sensing image data, meteorological data, and water quality data.

[0177] The multi-task learning module is configured to use a multi-task deep learning network to sequentially perform spatial feature extraction, serialization and global feature modeling on a multi-source data cube, and fuse the global feature representation obtained by modeling with the generated three-dimensional frequency domain feature vector to obtain a feature map.

[0178] The preliminary judgment module is configured to use a feature map to obtain a binary segmentation probability map of algal blooms using a segmentation head, and to obtain a continuous value distribution map of chlorophyll a concentration using a grading head.

[0179] The semantic extraction module is configured to generate image semantic embedding vectors based on global feature representation, using a semantic embedding head, and then semantically aligning the image semantic embedding vectors with a pre-trained knowledge graph to generate judgment information text.

[0180] The cross-validation module is configured to cross-validate the judgment information text, the binary segmentation probability map of algal bloom, and the continuous value distribution map of chlorophyll a concentration to obtain the final algal bloom risk judgment result.

[0181] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0182] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0183] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0184] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0185] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing intelligent identification method for algal bloom risk, characterized in that, Includes the following steps: Acquire remote sensing image data, meteorological data, and water quality data for the target area; The acquired remote sensing image data, meteorological data, and water quality data were preprocessed respectively; Frequency domain features are extracted from the preprocessed remote sensing image data. Gray-scale image blocks of a set band are selected and subjected to fast Fourier transform to obtain complex spectrum. Power spectral density is calculated, and the power spectral density spectrum is divided into frequency regions of different energy levels and the average energy is calculated to generate a three-dimensional frequency domain feature vector to characterize the spatial texture pattern of algal blooms. Construct a multi-source data cube, which stacks processed remote sensing image data, meteorological data, and water quality data; Using a multi-task deep learning network, spatial feature extraction, serialization, and global feature modeling are performed sequentially on a multi-source data cube. The global feature representation obtained from the modeling is fused with the generated three-dimensional frequency domain feature vector to obtain a feature map. Based on the feature map, a binary segmentation probability map of algal bloom is obtained using the segmentation head, and a continuous value distribution map of chlorophyll a concentration is obtained using the grading head. Based on global feature representation, image semantic embedding vectors are generated using a semantic embedding head. Pre-trained knowledge graphs are then used to semantically align the image semantic embedding vectors to generate judgment information text. Cross-validate the assessment information text, the binary segmentation probability map of algal bloom, and the continuous value distribution map of chlorophyll a concentration to obtain the final algal bloom risk assessment result.

2. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The preprocessing of the acquired remote sensing image data includes: radiometric calibration of the remote sensing image data, converting the digital quantization values ​​into atmospheric top radiance, atmospheric correction of the remote sensing image data to obtain the water body remote sensing reflectance, and calculating the normalized differential chlorophyll index and phytoplankton index within the water body area.

3. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The preprocessing of the acquired meteorological data includes: meteorological data such as sea surface temperature, wind speed, wind direction and precipitation; interpolating the meteorological data onto the same spatial grid as the remote sensing image; calculating the daily average or the cumulative value of the previous day that matches the transit time of the remote sensing image; and decomposing the wind speed vector into eastward and northward components.

4. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The preprocessing of the acquired water quality data includes: water quality data including chlorophyll a, total nitrogen and total phosphorus concentrations; outlier measurements are removed using the interquartile range method; if the density of monitoring stations is less than the set value, spatial distribution maps are generated using inverse distance weighting or kriging; the measurement value closest to the date of the remote sensing image is selected; if the dates do not match, linear interpolation is performed.

5. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The process of dividing the power spectral density spectrum into frequency regions of different energy levels and calculating the average energy includes: dividing the power spectral density spectrum into three sector regions of low frequency, mid frequency and high frequency in polar coordinates, and calculating the average energy value of each region. Among them, the low-frequency fan-shaped region is used to characterize the components in the image whose velocity changes are less than a set value, corresponding to large-area, uniformly distributed algal blooms; The mid-frequency fan-shaped region is used to characterize the components of image velocity changes within a set range, corresponding to algal blooms with filamentous, strip, and blocky structures; High-frequency sector regions are used to characterize components in an image whose velocity changes are greater than a predetermined value, corresponding to fragmented distribution, algal bloom boundary contours, or noise. Based on the average energy values ​​calculated for each sector region, a three-dimensional frequency domain feature vector is constructed as a quantitative description of the spatial distribution pattern of algal blooms.

6. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The process of constructing a multi-source data cube includes: the shape of the multi-source data cube is [Batch_size, Channels, Height, Width], where Batch_size is the number of samples input into the model for processing at one time, Channels are the feature dimensions of remote sensing image data, meteorological data and water quality data, Height is the spatial height dimension of the cropped remote sensing image, and Width is the spatial width dimension of the cropped remote sensing image.

7. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The multi-task deep learning network includes an encoder, which is used to perform preliminary spatial feature extraction on a multi-source data cube. The extracted feature map is divided into several fixed-size image blocks according to spatial location. The features of each image block are flattened and mapped to a unified dimension through linear projection to form a set of serialized feature vectors. A learnable classification token feature is added to the front of the serialized feature vectors to aggregate global information. The serialized feature vectors after the addition are used to model global features through a multi-head self-attention mechanism to obtain a representation with global dependencies. The three-dimensional frequency domain feature vector is projected onto the same dimension as the global feature through a fully connected layer and then fused with the classification token feature element-wise to enhance its feature representation.

8. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The multi-task deep learning network includes a decoder, which includes a segmentation head, a hierarchical head, and a semantic embedding head. The segmentation head is used to receive the feature map output by the encoder and output a binary segmentation probability map of algal blooms through convolution and Sigmoid activation function to characterize the spatial distribution range and boundary of algal blooms, thereby supporting the drawing of algal bloom distribution maps and area calculation. The grading head is used to receive the feature map output by the encoder, and after convolution and linear activation function, it obtains the continuous value distribution map of chlorophyll a concentration, and discretizes it into a risk level map for algal bloom level classification and risk warning. The semantic embedding head takes the global semantic vector output by the encoder as input, and generates an image semantic embedding vector through projection transformation. The image semantic embedding vector encodes the overall algal bloom status of the image patch and its environmental context information, and is aligned with the structured semantic tags and ecological knowledge graph to realize the semantic parsing of remote sensing image results.

9. The remote sensing intelligent identification method for algal bloom risk as described in claim 1, characterized in that, The training process for a knowledge graph includes: A structured semantic tag library was constructed to generate structured tags for each training sample, including algal bloom state, dominant species, density level, and environmental conditions, and then converted into natural language description sequences. The text description is encoded into a feature vector using a large language model. The model is trained using a loss function to bring matching image-text pairs closer together and push away mismatched pairs, thus achieving cross-modal semantic alignment. For each structured label, a corresponding text embedding vector is output using the trained large language model. Based on the structured label and the text embedding vector, a knowledge graph is formed.

10. A remote sensing intelligent identification system for algal bloom risk, characterized in that, include: The data acquisition module is configured to acquire remote sensing image data, meteorological data, and water quality data of the target area; The preprocessing module is configured to preprocess the acquired remote sensing image data, meteorological data, and water quality data respectively. The feature extraction module is configured to extract frequency domain features from the preprocessed remote sensing image data, select grayscale image blocks of a set band, perform fast Fourier transform to obtain complex spectrum, calculate power spectral density, divide the power spectral density spectrum into frequency regions of different energy levels and calculate average energy, and generate a three-dimensional frequency domain feature vector to characterize the spatial texture pattern of algal blooms. The data stacking module is configured to construct a multi-source data cube, which stacks processed remote sensing image data, meteorological data, and water quality data. The multi-task learning module is configured to use a multi-task deep learning network to sequentially perform spatial feature extraction, serialization and global feature modeling on a multi-source data cube, and fuse the global feature representation obtained by modeling with the generated three-dimensional frequency domain feature vector to obtain a feature map. The preliminary judgment module is configured to use a feature map to obtain a binary segmentation probability map of algal blooms using a segmentation head, and to obtain a continuous value distribution map of chlorophyll a concentration using a grading head. The semantic extraction module is configured to generate image semantic embedding vectors based on global feature representation, using a semantic embedding head, and then semantically aligning the image semantic embedding vectors with a pre-trained knowledge graph to generate judgment information text. The cross-validation module is configured to cross-validate the judgment information text, the binary segmentation probability map of algal bloom, and the continuous value distribution map of chlorophyll a concentration to obtain the final algal bloom risk judgment result.

Images